Horizons of AI: Ethical Considerations and Interdisciplinary Engagements

Artificial Intelligence raises concerns, especially in areas where human beings are affected by its decisions. Trust, explainability, and accountability are therefore crucial for legal compliance and to increase end-user confidence. This is particularly relevant when Intelligent Controls are employed in Swarm Robotic systems, which consist of numerous independent units functioning as a collective. Here, the complexity of trustworthy Artificial Intelligence solutions may escalate rapidly, necessitating the development of swarm intelligence paradigms that differ from conventional methods, used for single intelligent agents. Interactions between humans and Artificial Intelligence raise critical questions: at the internal system level, where human-in-the-loop control is often required, and at the external level, where intelligent agents’ actions may impact humans, potentially leading to ethical dilemmas. This paper proposes a new extension to Human Swarm Teaming-3 architecture adding explainability and accountability through Distributed Autonomous Organization and Distributed Ledger Technologies. Finally, an implementation of this new Human Swarm Teaming-4 architecture is proposed using a Robot simulator software and Distributed Ledger testnet.

The problem of coordinating the operations of a robot team comprising unmanned ground vehicles and unmanned aerial vehicles, toward a single leader formation, is studied in the Ramadge-Wonham framework. In this robot team, the members of a specific group of unmanned vehicles (UVs) can take over both roles of leader and follower, as they have the ability to exchange roles during a task. The rest UVs of the robot team can have only the role of the follower. In the present paper, the modeling of the robot team is parametric with respect to the total number of the team’s UVs, as well as the number of UVs that can take over both roles. A supervisor, distinguishing the roles of leader and follower, is developed for UVs being able to perform both roles. Several rules, providing desired behavior of each UV, with respect to formations, are translated to a set of corresponding regular languages. Each language is realized by an appropriate two-state supervisor automaton, thus facilitating the implementation of the designed controllers and providing low computational complexity. The property of physical realizability of the supervisors and the property of nonblocking of the total controlled automaton are proved.

Traffic monitoring can be supported by stationary or dynamic camera systems. Whereas the analysis of stationary camera data has a long tradition, the analysis of dynamic sensors (recorded by Unmanned Aerial Systems (UAS) or dashcams in cars or trucks) has gained interest in the recent past. This paper proposes a labeling enhancement of a dashcam video repository and the evaluation of several recent deep learning algorithm (LSTM, Inception V3, VGG16, MobileNet V2, etc.) for classifying traffic accidents in dynamic camera systems. Based on fine-tuning methods and data preprocessing steps optimized results for the individual deep learning approaches could be achieved and an overall analysis is presented.

Using a 2-D Convolutional Neural Network (CNN) method for fault diagnosis, appropriate fault diagnosers are composed, for a type of Automated Guided Vehicles (AGVs), having four categories of actuators and three categories of sensors. The actuators and the sensors are modeled as Discrete Event Systems (DESs) in the Ramadge-Wonham (RW) framework. The diagnoser consists of an online level for fault detection and isolation and an offline level for the CNN training. The model of the presence and alter of faults, in AGV’s actuators and sensors, is expressed in the form of appropriate two-state automata that incorporate corresponding events produced by the fault diagnosers. A set of safety rules, considering the eventual presence of faults, will be introduced. The safety rules will be realized in the form of supervisor automata. The marked behavior of the controlled automata of the actuators and sensors will be provided to validate the satisfactory performance of the proposed method.

This article examines the use of the “ethics by design” technique, as it has been evaluated and introduced by the European Commission, in the development of AI systems used for decision-making purposes by European civil or administrative justice, in order to address the basic challenges that these systems may pose to the effective and lawful judicial procedure. The study focuses on the fundamental principles that should govern the ethical development of AI systems and in what ways these principles can be transformed in internal values of the design phase of the system used in the judicial sector, and incorporated into its lifecycle operation. Finally, we explore the limitations of the use of this technique in building the trustworthiness of automated decision-making systems in the judicial procedure, arguing in favour of a human-centred decision-making process.

This study explores the perception of bias in AI-generated news articles among Generation X and Generation Z. The primary objective is to examine whether the generational differences of age affect the perception of bias and the perceived usefulness of AI-generated content. A quantitative approach is employed, utilizing a survey with 60 participants equally divided between the two generations. The results indicate that Generation X perceives AI-generated news as more useful and less biased compared to Generation Z, although these differences are not statistically significant. Additionally, a significant positive correlation is found between perceived usefulness and perceived bias, suggesting that participants who found AI-generated news useful also perceived less bias in the content. The findings highlight the importance of considering generational perspectives when implementing AI in news production. Media and technology companies can leverage these insights to refine their strategies and foster user trust and acceptance. The study’s limitations include a small sample size and focus on German-speaking participants, limiting the generalizability of the results. Future research should include larger, more diverse samples and incorporate qualitative methods to gain deeper insights into individual perceptions and attitudes toward AI-generated news. Overall, this study provides valuable insights into the perception of bias and acceptance of AI in news production, laying the groundwork for further investigations and practical applications in this field.

The advent of Artificial Intelligence (AI) technologies has brought significant advances and efficiencies, making it a pervasive presence in our daily lives. The application of AI is exploding in healthcare, finance, and transportation, as well as in education. However, the integration of AI in education is not without challenges and risks. In this context, the role of research and researchers becomes of paramount importance.After outlining what AI ethics in education is concerned with and the role of language choice in communication, starting from a systematic review focused on the ethical challenges of AI in Higher Education (HE), a collateral study is conducted obtaining a selection of materials analyzed through Voyant Tools. In the 25 texts, five words emerged as foundational elements of this research’s field: “AI,” “students,” “education,” “learning,” and “ethical.” The result promoted the subsequent investigation concerning how the words relate in the linguistic context. A general alignment of the research community emerges, and a balance between action, promotion, innovation, prudence, and reflection is proposed. This study addresses a request to take care of the ethical dimension, asking the research world to give it even more space in the research papers and considering it fundamental to create safe, responsible, and ethical teaching and learning practices. Aware of the era in which we live, the study prompts procedures and questions and opens avenues for future research, asking it to delve into how the research world is addressing the ethical dimension of AI in HE.

This paper examines the impact of Artificial Intelligence (AI) on research ethics, exploring both the challenges and opportunities presented by AI in academic contexts. We investigate AI as a subject of research ethics, analyzing the challenges faced by Institutional Review Boards in evaluating AI research proposals. The paper discusses the influence of AI on data collection and analysis, discussing privacy concerns and the implications of regulations such as GDPR and the EU AI Act. We explore the ethical dimensions of AI-generated content in academic research, including literature reviews and meta-analyses, as well as in peer review and publication processes. Finally, we discuss future directions and regulatory challenges, emphasizing the need for adaptive ethical frameworks and highlighting the balance between leveraging AI’s potential to enhance research capabilities and the need to uphold ethical standards and human values in academic inquiry.

The escalating complexity of modern data landscapes necessitates innovative approaches to streamline Extract, Transform, Load (ETL) processes. This research investigates the potential of Large Language Models (LLMs), specifically BERT and T5, to automate metadata extraction from tabular data exclusively based on data values, bypassing the need for traditional header information. Our study evaluates the ability of pre-trained LLMs to accurately predict schema elements such as data types, constraints, formats, and relationships with minimal training data. Using the diverse AdventureWorks datasets, we rigorously test the LLMs capacity to handle real-world ETL challenges. Our findings demonstrate that LLMs can effectively predict metadata with high accuracy, significantly reducing manual effort and accelerating ETL pipelines. By integrating LLMs into ETL workflows, we achieve substantial improvements in efficiency, timeliness, and accuracy. This research underscores the transformative potential of LLMs in revolutionizing data engineering practices and provides a foundation for future explorations in this domain.

Interpretability can be defined as the ability to understand a neural network and explain its behavior. In practice, Multi-Layer Perceptrons (MLPs) are considered opaque due to their size and complexity. To overcome this, a more interpretable alternative was proposed recently, i.e. Kolmogorov–Arnold Networks (KANs). In this work, we discuss the difference between Kolmogorov–Arnold representation Theorem (KAT) and Universal Approximation Theorem (UAT), i.e. the underlying theorems of KAN and MLP. Moreover, we highlight that KAN violates the size constraint in KAT; thus, it is not more interpretable than MLP. Furthermore, we present practical scenarios and argue that KANs are less interpretable than MLPs when dealing with high-dimensional data like images and text. Additionally, we perform several experiments and show that KANs contain larger number of trainable parameters compared to MLPs, causing a significant additional computational cost without noticeable performance improvement.

Dementia, a widespread neurodegenerative condition, presents significant challenges in early diagnosis and intervention. This study investigates innovative methods for detecting dementia by analyzing linguistic patterns in speech transcripts through advanced machine learning techniques. Using datasets from DementiaBank, including the Pitt Corpus and ADReSS challenge datasets, we leveraged Large Language Models (LLMs) to assess cognitive and linguistic features. Our methodology encompassed comprehensive data preprocessing, feature extraction from the ‘Cookie Theft’ picture description test, and model fine-tuning. We explored various transfer learning strategies with pre-trained models such as BERT, DistilBERT, RoBERTa, Mistral, and Llama. Our findings indicate that LLMs, particularly when optimized with Low-Rank Adaptation (LoRA) and Quantization (QLoRA), can achieve a dementia detection accuracy of 96%.

Cognitive impairments in attention are prevalent among patients with Alzheimer’s disease (AD), Parkinson’s disease (PD), dementia with Lewy bodies (DLB), and Parkinson’s disease dementia (PDD). Electroencephalograms (EEGs), particularly event-related potentials (ERPs), provide valuable insights into these impairments. This study explores the use of Hjorth descriptors—activity, mobility, and complexity—from ERP signals to automatically classify these cognitive disorders. We analyzed EEG data from five subject groups (DLB, PD, PDD, AD, and healthy controls) using k-nearest neighbors, random forest, and gradient boosting classifiers. Data were collected from 90 subjects and each participant underwent neuropsychological assessments and EEG recordings during auditory oddball-distractor tasks. ERP segments were made based on three different events used during the paradigm. Our findings revealed significant differences in the Hjorth descriptors when compared with healthy controls, particularly in the PDD group, where decreased mobility indicated lower mental activity or alertness. The Random Forest classifier outperformed other methods, emphasizing its potential for effective differentiation of cognitive disorders. This study highlights the utility of EEG and machine learning in the early detection and classification of neurodegenerative diseases, offering valuable insights for better patient management.

In a landscape characterized by multifaceted uncertainties, borrowers face challenges in meeting their loan obligations while financial institutions grapple with accurately assessing credit risk. To address these challenges, financial institutions are increasingly integrating environmental, social and governance (ESG) criteria into their risk management strategies. Concurrently, advancements in artificial intelligence (AI) and machine learning (ML) are revolutionizing the current credit risk models. By utilizing bank-level data for the top twenty (20) European banks financing maritime borrowers, for the period 2010–2022, we investigate the significance of integrating ESG criteria in predicting shipping loan defaults. By applying the ML approach, our results demonstrate that the integration of the ESG criteria improves the predictive capability of the ML models concerning maritime loan defaults. Notably, among the ESG criteria, the combined ESG score offers the best predictive accuracy, followed by the environmental, governance and social criteria.

Organizations such as hotels can utilize AI to analyze reviews cost-effectively, thereby eliminating the need for customer service experts. Deep learning (DL) methods, while accurate with large datasets, often lack explainability. This paper utilizes a Bi-LSTM for sentiment analysis on a collection of 515,000 hotel reviews, achieving an accuracy of 89.67%. We then apply both SHAP and LIME, two prominent but different XAI tools, to explain our model. These techniques clarify how specific words impact sentiment, validating and interpreting the Bi-LSTM model’s decisions.

Context. As software systems become increasingly intertwined with societal infrastructure, the responsibility of software professionals to ensure compliance with non-functional requirements, including but not limited to safety, privacy and non-discrimination, is paramount. Motivation. Ensuring fairness in pricing algorithms allows for fair access to essential services by not discriminating on protected attributes. Method. We replicated a previous empirical study that used black box testing to audit pricing algorithms used by Italian car insurance companies, accessible through a popular online comparator website. In comparison with the aforementioned study, we augmented the number of tests and the number of demographic variables under analysis. Results. The present study corroborates and extends previous findings, highlighting the persistent nature of discrimination across time. Demographic variables significantly impact pricing, with birthplace remaining the primary discriminatory factor against individuals not born in Italian cities. Furthermore, the analysis revealed that driver profiles can determine the number of quotes available to users, thereby denying equal opportunities to all. Conclusion. The study emphasises the significance of incorporating non-discrimination testing into software systems that have a direct impact on individuals’ daily lives. The analysis of algorithms over an extended period facilitates the assessment of their evolution over time. Furthermore, it illustrates the potential of empirical software engineering to enhance the accountability of software systems.

Can we have better understanding for bank customers through social media? Social media has faced exponential growth during the last few years, creating a very rich source of information and insights on every aspect more than ever. The customers of mainstream lenders (i.e., banks, credit unions), after all, are normal people who interact on social networks and have connections, opinions, sentiments, and reactions in different contexts. Banking credit score calculations consider various financial parameters, overlooking other personal factors of a bank customer that might have a great impact on his financial abilities. Analyzing social media interactions of a bank customer would give an additional layer of eligibility evaluation to the bank customer, which should influence the overall credit scoring calculations process. In this paper, we propose the Social-based Credit Scoring (Socio-CS) approach as a new method for credit score calculation using machine learning techniques, in which new parameters derived from the social media data are integrated into the financial criteria to provide a global representation of the potential borrower and thus improve the trust level of credit scores. A bank dataset is integrated to a twitter dataset to build a case study for experimentation, considering sentiment analysis. The results show that Random Forest achieves the highest average accuracy of 81%, whereas Logistic Regression and Support Vector Machine achieve 72% and 59%, respectively.

The current technology in hydraulic systems for production processes typically relies on analog spool valves, such as analog and servo valves. However, these valves cause high power losses due to significant pressure drops during valve control. The emerging trend is to replace flow control valves with robust and cost-effective digital on/off valves, thereby reducing power losses and enhancing the overall efficiency of hydraulic systems providing hybrid (digital and analog) model descriptions. In this paper, a fuzzy control scheme is designed to achieve precise position control of a digital hydraulic actuator, including a set of 16 on/off hydraulic valves. The fuzzy controller employs a type 1 Takagi–Sugeno fuzzy inference system (FIS) feeding the performance error and its derivative. The controller's signal activates the appropriate set of on/off valves. Simulation results demonstrate the excellent performance of the closed-loop system.

With the evolution of the Internet of Things (IoT), a vast amount of data is generated from various intelligent devices and sensors. Data streaming in IoT systems may face quality issues like incompleteness due to sensor failures, network disruptions, or transmission errors. Addressing this problem is crucial, as unhandled missing data leads to inaccurate and unreliable analysis, compromising decision-making processes. This paper presents a multi-layer, data-driven methodology for imputing univariate IoT data streams with significant missing gaps by combining statistical analysis, signal processing, and Deep Learning algorithms. We propose a novel approach called the ES-CC-BiLSTM model, which consists of four phases: first, inserting missing data using the Exponential Smoothing technique; second, generating diverse statistical and temporal features; third, conducting cross-correlation analysis to identify the most critical features; and finally, utilizing the Bidirectional Long Short-Term Memory (BiLSTM) model to impute missing data. The model is evaluated using real-world water consumption datasets with artificially simulated missing values ranging from 10 to 50% and long interval gaps of up to 48 consecutive missing values. Experimental results demonstrate that the proposed framework reduces imputation error by up to 30% compared to the second best-performing model, and significantly outperforms other statistical and Machine Learning methods in terms of RMSE and MAE metrics.

In the present paper, using input-output models of manufacturing machines in Discrete Event System (DES) form, flexibility of a manufacturing floor adopting two scenarios is studied. The manufacturing floor has a parametric number of machines, while the two architectures are the open chain architecture and the ring architecture. Developing a parametric number of embedded supervisors, based on the cooperative pair approach, and intelligent machine vision system, dedicated to product inspection, as well as by developing two switching supervisors, the two-case flexibility of the manufacturing floor is accomplished. This way, a two-level hierarchical supervisor control scheme is proposed. Finally, implementation of the supervisors in CODESYS is presented.

In this work, a novel approach to earthquake detection by integrating deep learning architectures with decentralized data management is introduced. To this end, variational autoencoders are trained within the OASEES framework, employing the InterPlanetary File System for data and model storage, moving beyond traditional centralized cloud/edge processing. The obtained experimental results demonstrate the model’s ability in classifying seismic events, achieving an accuracy level of $$97.24\%$$ 97.24 % . The proposed distributed architecture not only achieves top performance, but also aligns with the heterogeneous cloud-fog-edge computing continuum, offering improved data governance and control.

Semantic segmentation models are widely used in various computer vision tasks. However, these models often suffer from biases and inefficiencies, which can limit their performance. In this paper, we propose a novel approach of model pruning using explainable AI (XAI) techniques. The proposed approach aims at identifying and eliminating non-pertinent channels in convolutional layers. It is applied for segmentation-based models, where we use XAI as the criterion for pruning techniques specifically tailored for segmentation tasks. By guiding the pruning process by means of explainability metrics, thorough experiments were conducted on the DIVA-HisDB benchmark dataset to correct model biases and enhance overall efficiency and performance.

In this work, we present a novel web ontology of cultural conceptualizations that is based on the Cultural Linguistics framework developed by Farzad Sharifian. This theoretical-analytical framework explores the relationship between Cognition, Culture and Language (CCL), focusing on three components, i.e. cultural schemas, categories and metaphors, through which linguistic expressions and their cognitive underpinnings are analysed. The CCL ontology reuses the DOLCE (DOLCE+DnS Ultralite) foundational ontology, selected for its cognitive emphasis and extends it with domain-specific concepts and relationships. Potential use in comparative cultural studies and cross-cultural educational contexts is demonstrated through a set of use cases constructed from published research on political discourse in Ghana and China and the comprehension of Australian Aboriginal English narratives.

In the present paper, we exploit the results of a recent work on multisubject book classification by extending its application to book collections written in languages other than English, specifically in Greek. The proposed classification method consists of utilizing the word statistics in the book’s Table of Content as well as in a controlled subject-term vocabulary, in combination with the Latent Dirichlet Allocation (LDA), a well-known machine learning technique for discovering hidden topics in a corpus of documents. The proposed method was theoretically formulated and validated through an extensive set of experiments performed on Springer’s English language e-book collection. Now, the classification method is applied on book collections written in Greek: a set of about fifty thousand academic books, provided by commercial publishers through the EVDOXUS service, and a more limited collection of digital books publicly available with open licenses (the KALLIPOS collection). The derived qualitative and quantitative results show the language-neutral applicability of the proposed approach, with the Latent Dirichlet Allocation method, combined with simple Bayesian inference, also being highly effective in analysing Greek language collections. Upon examining traditional metrics such as precision and recall, it is evident that their values converge and surpass a score of 0.82 when classifying unknown documents in Greek across 26 different subjects. This confirms the efficacy of the suggested approach and paves the way for the application of the proposed classification method to multilingual collections, provided that the vocabulary of the subject terms is available in other languages of interest. The availability of common Natural Language Processing tools, as for example stemmers, lemmatizers, common-word filters, required for document preprocessing, is taken for granted in all modern Natural Language Processing programming platforms.

We present an open-source presentation video retrieval system, which performs automatic video content analysis of a presentation video (e.g., a lecture or talk) using state-of-the-art transformers for OCR and ASR and a specific shot detection approach. It automatically recognizes text and speech in the video and indexes it with the corresponding keyframe, enabling content-based retrieval. By combining OCR and ASR with a simplistic, efficient UI, we provide an effective and easy-to-use solution to retrieve specific video segments from large collections of lecture videos. Users can query for specific video segments displayed by representative keyframes. In this paper, we compare our system to related work and give an in-depth description of the functionalities and frameworks used to build the system. Additionally, we discuss the systems’ architecture and present a preliminary evaluation of our system, which shows promising results for different types of presentation videos. The entire system is released as an open-source tool and can be freely used by all colleagues who want to make their presentation videos searchable.

This research presents a novel approach to automating the disassembly process of personal computers (PCs) as a critical step toward circular manufacturing of Waste Electrical and Electronic Equipment (WEEE). By integrating artificial intelligence (AI) and robotics, we aim to optimize resource recovery and minimize environmental impact. The study focuses on developing a system that accurately identifies and localizes key PC components, such as motherboards, processors, cooling fans, and screws, using modern computer vision techniques. YOLO object detection algorithm is employed for 2D image analysis, while morphological computer vision is utilized to transform 2D coordinates into 3D locations. A collaborative robot is then tasked with unmounting the identified components based on the generated 3D locations. This research contributes to the Horizon Europe’s Circular TwAIn project by demonstrating the potential of AI and robotics to revolutionize WEEE management and promote a more sustainable electronic waste lifecycle in Europe.

The problem of controlling a differential drive mobile robot carrying different objects with different and known mass and inertia is investigated. The design objective is noninteracting control of robot’s velocity and orientation angle, while ensuring stability, asymptotic command following, and model following for all object loads. In order to satisfy this objective, a common dual-layer PI-PID controller is proposed. The inner layer control consists of two PI controllers dedicated to regulating the angular velocities of the active wheels. For the outer layer, a multivariable PID controller is designed based on the inner closed-loop system. The parameters of both control layers are optimized using a metaheuristic algorithm that minimizes specific cost functions, representing the deviation of the linear closed-loop systems (corresponding to different loads) from an ideal model. The effectiveness of the proposed control scheme is demonstrated through simulations of the application of the control scheme to the original nonlinear system models.

The European industry twin green and digital transition is critical to foster sustainable manufacturing practices that minimize and prevent the environmental impact and resource exhaustion, while ensuring European global competitiveness. AI can play a pivotal role in this transitioning helping to optimize resources, enhancing efficiency, and reducing waste. Nonetheless, AI adoption in the industry is still limited due to the skills gap that hinders the management of these advanced systems. This paper proposes a self-adaptive pipeline for ML with the incorporation of a new Overfitting Index (OI) for self-parameter tuning emphasizing overfitting prevention. The pipeline incorporates self-parameter exploration capabilities exploiting surrogate models to improve computational efficiency. The proposed OI for ML function is evaluated in a well-known regression problem, prior to its real evaluation in an aluminum recycling application to support operators selecting the best combination of scraps. Results indicate that configurations with lower OIs demonstrate superior generalization and robustness, with the surrogate model effectively identifying and refining high OI configurations. The proposed methodology provides a baseline for future development and integration of self-adaptive and self-improving ML solutions in the industry.

In the present paper, an event-driven supervisor controller of a robotic cell for palletizing is designed and implemented. The cell consists of an M-410iB/450 Fanuc Robotic Manipulator and three conveyor belts. The task of the Robotic Manipulator is to transfer the products from belt to belt. The cell is modeled using Discrete Event Systems. The sensors and actuators of the conveyor belts are connected to a Programmable Logic Controller and are studied using Input–Output (I/O) Discrete Event System models. The supervisor design is based on the Supervisory Control Theory. A modular supervisor control architecture is designed. Regarding robotic manipulators, a parametric supervisor is proposed to accomplish trajectory following. Also, a set of supervisors is proposed for the safe functionality of the gripper. Finally, for the coordination of the devices of the cell a set of supervisors is designed, based on corresponding regular languages that describe the desired behavior. The robotic cell control programming is implemented and simulated in the ROBOGUIDE Simulation Software.

In this study, a low-power analog integrated classifier for biomedical applications based on an ANN is introduced. The proposed architecture consists of four analog circuits as building blocks. All circuits operate in sub-threshold region in order to achieve low-power consumption. Post-layout simulations using the Cadence IC Suite and the TSMC 90 nm CMOS technology demonstrate that the proposed analog classifier operates properly with good sensitivity. The implemented classifier is trained using software and it is compared with related classifiers. It can be used as a building block for biomedical monitoring.

Recent advances in information retrieval introduced Large Language Models (LLM) as chatbot assistants. However, LLM’s reliance on implicit knowledge alone is prohibitive for knowledge-intensive tasks. The retrieval-augmented generation (RAG) technique extends the capabilities of LLMs by supplementing contexts to generate contextually relevant responses. RAG relies on a retriever model that selects documents from a knowledge database. In niche industries with limited training data, generic retriever models often underperform, leading to incorrect chatbot responses. This paper aims to improve a drilling text document retriever by adapting a generic dense embedding model. Using a contrastive learning technique, we aim to maximize the similarity scores of relevant items while pushing dissimilar items apart. We train and evaluate the models based on retrieval accuracy using the Norwegian Petroleum Directorate data. Our results show that contrastive learning enabled the adapted model to retrieve 94% more relevant information than its generic version. The adapted model also outperformed the best commercial model by 80% despite having 6x fewer dimensions. Moreover, based on our sensitivity analysis, the different hyperparameter configurations resulted in minimal variations in the adapted model performance. We release our generated datasets for the public to facilitate further research.

Interpretation and visualization of the behavior of detection transformers often highlight image areas attended to by the model but provide limited insight into the semantics that the model is focusing on. This paper introduces an extension to detection transformers that learns and utilizes prototypical local features for object detection, termed prototypical parts. These features are designed to be mutually exclusive and align with the detection classes of the model. The proposed extension consists of a bottleneck module, the prototype neck that computes a sparse representation in terms of prototype activations, and a novel loss term that aligns prototypes with object classes. This approach leads to interpretable representations in the prototype neck, allowing visual inspection of the image content as perceived by the model and a better understanding of model reliability. Results show that our method incurs only a limited performance penalty. Furthermore, we provide examples that showcase the explanatory power of our approach, justifying the slight performance reduction.

This paper presents a comparative analysis of building segmentation using pre-trained deep learning models on different data types: RGB images and LiDAR data. The study utilizes the MapAI dataset and evaluates the performance of the Segment Everything Everywhere All at Once (SEEM) model for direct inference and the You Only Look Once (YOLOv8) model fine-tuned for building segmentation. Performance is assessed based on Intersection over Union (IoU) and Boundary Intersection over Union (BIoU) metrics. Our findings reveal that SEEM inference on RGB images yields better results compared to LiDAR data, and a similar trend is observed for the YOLOv8 model. Additionally, in the context of fine-tuning, the YOLOv8 model significantly outperforms the trained shallow model and trained weighted U-Net ensemble on RGB images, demonstrating its superior capability for precise building segmentation. Conversely, for LiDAR images, the customized U-Net model yields better results than YOLOv8, highlighting the importance of model selection based on the data type. This comprehensive evaluation underscores the critical role of model adaptation and data characteristics in remote sensing applications for building extraction.

Complex signals are ubiquitous in various fields of engineering and science. The complex Kalman filters are used to process complex signals. The complex Kalman filter is associated with linear models in linear estimation, while the complex extended Kalman filter is associated with augmented models or widely linear models in nonlinear estimation. Both complex Kalman filter and complex extended Kalman filter utilize the Kalman filter gain in order to compute the estimation and the prediction of the n-dimensional state using the m-dimensional measurement. In this work, variations of complex Kalman filter and of complex extended Kalman filter are derived; the proposed algorithms eliminate the Kalman filter gain. The proposed Kalman filter with gain elimination may be faster than the traditional Kalman filter; in fact the model dimensions determine the fastest filter. Also, steady-state complex Kalman filters are derived, which require the solution of the complex Riccati equation. Per step and doubling iterative algorithms for the solution of the complex Riccati equation are presented.

The digitalisation of services associated with Autism is considered important although its implementation is limited yet. A Digital Twin (DT) is a virtual representation of a physical entity, which not only exhibits any changes in the physical entity, but can affect it in return, as well. This technology holds promises to enhancing autism research, particularly autism assessment, intervention, and rehabilitation. However, despite its growing application in healthcare over the last decade, it has not been broadly adopted in autism research yet. This study aims to present the design of a DT and introduce it in autism educational/intervention processes as a digital copy of them. The development of such a DT can result in real-time performance and supervision of the educational/intervention processes without an educator’s/caretaker’s involvement. The study also discusses potential limitations encountered during the DT design process, as well as recommendations for future research, regarding its effectiveness and applicability.

The expanding field of psychosocial online email counselling has created a demand for advanced tools to support counsellors in managing diverse client interactions. This paper introduces CAIA (Counsellor Artificial Intelligence Assistant), a comprehensive AI-driven system currently under active development, designed to condense and present information to counsellors in a structured manner. CAIA integrates several AI-based features, including Automatic Subject Generation, Case Summarisation, Timeline Management, Social Network Graph creation, and Case Classification. The features are designed to assist the counsellors, allowing them to engage more deeply with each case while ensuring that the client’s needs remain paramount. Each feature presented in the paper addresses a specific problem in psychosocial online counselling, proposes a solution, discusses a technical implementation approach, and highlights the challenges encountered. Additionally, the paper discusses ethical considerations associated with an AI integration.

This research presents a novel, low-power, fully analog integrated circuit for classifying fetal health data. The circuit leverages a centroid-based approach, employing Vector Length calculations and current comparisons. Notably, it achieves high classification accuracy $$(93.8\%)$$ ( 93.8 % ) while consuming only 675 nW. This design prioritizes energy efficiency by operating transistors in the sub-threshold region while maintaining robustness. To validate its effectiveness, the proposed circuit is compared against existing software and hardware-based classifiers. Python programming language was used to train the classification model and process the resulting data. The design and simulations were conducted within the Cadence IC Suite environment, utilizing the TSMC 90nm CMOS process technology.